object-oriented classification of alpine landforms from an ... · a digital elevation model (dem)...

Grazer Schriften der Geographie und Raumforschung Band 45/ 2010

53

Object-oriented classification of alpine landforms from an ASTER scene and digital elevation data (Reintal, Bavarian Alps)

pp. 53 - 62

Abstract

High mountain regions represent diffi cult terrain for detecting rock and sediment storage areas. By means of a satel-lite scene by the Advanced Spaceborne Thermal Emission and Refl ection Radiometer (ASTER) and a digital elevation model, the geomorphological setting of the Reintal subcatchment (17 km2) east of the Zugspitze is analysed. Characte-ristic landforms are classifi ed in an object-oriented approach comprising four spatial levels of differentiation. The complex, object-based decision tree hierarchy largely founds on fuzzy membership functions and to a lesser extent on a minimum distance classifi er. The fi nal landform classifi cation scores high in the accuracy assessments. The re-sults show that an identifi cation of the present-day pattern of geomorphological process units is possible by remote sensing. Besides, the approach provides a fi rst insight into the otherwise inaccessible upper regions of the study area which could not be included in any previous survey.

KEY WORDS: geomorphological landform detection, image segmentation, fuzzy logic, object-oriented classifi cation, Reintal, Northern Calcareous Alps

N. J. Schneevoigt1, S. van der Linden2, T. Kellenberger3, A. Kääb1, and L. Schrott4

1 Department of Geosciences, University of Oslo2 Geomatics Lab, Humboldt-University Berlin, Germany3 Remote Sensing Laboratories - Department of Geography, University of Zurich4 Department of Geography and Geology, University of Salzburg

54

geo10th International Symposium on High Mountain Remote Sensing Cartography

1. Introduction

High mountain regions display a ”geomorphic environ-ment of considerable diversity. This variability in both time and space is perhaps the single most signifi cant geomor-phic characteristic of the alpine zone” (Caine 1974). There-fore these fragile environments react very quickly and sen-sitively to global change (Kääb 2002). However, scientifi c knowledge about their geomorphologic process structure remains sketchy and incomplete, especially quantitatively. Similarly, the question of potentially mobilisable sedi-ments in the upper regions of high mountain catchments still calls for an answer (Schrott et al. 2003). Within a set of projects called ”Sediment Cascades in Alpine Geosystems” (SEDAG), the universities of Eichstätt, Erlangen, Halle and Bonn/Salzburg have developed a model to describe land-form evolution in high mountain regions. This study forms part of the SEDAG research and repre-sents the third of a series of papers: geomorphic systems theory and object-oriented remote sensing have been lin-ked in Schneevoigt and Schrott (2006) in order to convey the theoretical and conceptual background of the analy-sis. Schneevoigt et al. (2008) accentuates its geomorpho-logical side, particularly stressing the nature of the alpine landforms examined. In contrast, this paper at hand de-scribes the remote sensing methods employed in further depth.It aims at a semi-automatic classifi cation scheme for ge-omorphological landforms, which can supply otherwise inaccessible information besides assisting landscape mo-nitoring and mapping. As upper areas mostly cannot be observed from the ground, remote sensing applications represent a means of closing this gap which hampers a full understanding of the alpine sediment cascade. Many studies on high mountain geomorphology use GIS cou-pled with remote sensing data, whereas only few em-ploy genuine remote sensing techniques (for details see Schneevoigt et al. 2008, Schneevoigt and Schrott 2006). Object-oriented image segmentation prior to classifi ca-tion constitutes a novel and promising approach (Blasch-ke et al. 2002, Benz et al. 2004). This relatively new trend has also found its way into high mountain applications, e.g. with Giles and Franklin (1998) classifying geomorpho-logical slope units.

2. Geographical setting

The Reintal valley is situated 7 km south of the town of Garmisch-Partenkirchen in the Bavarian Alps (Fig. 1). It ex-tends over 8 km in predominantly dolomitised limestone or Wettersteinkalk. As the Zugspitzplatt is not linked to the valley in terms of sediment transfer, it has been exclu-ded from the study area amounting to 17 km2. No glaciers persist today, but Pleistocene glaciations have typically

shaped cirques and hanging valleys in the upper regions, over steepened rockwalls and a broadened valley bot-tom. The relative relief within the study area amounts to 1690 m reaching a maximum of 2744 m asl. at Hochwan-ner peak, a fact which amongst others confi rms the high mountainous nature of the Reintal.Today, 79% of the sediment stores on the valley fl oor are relict or inactive and completely decoupled from the sedi-ment cascade system. Avalanche and debris fl ow tracks, alluvial fans and fl oodplains represent the most active storage types. In general, process activity rises with valley altitude. Very low clastic sediment output turns the Rein-tal into an effective serial sediment trap. As most active landforms receive input only, sediment stores build up quickly (Schrott et al. 2003).

3. Data basis

In this study, ASTER scenes were assessed because of their spatial resolution, pricing and near global coverage (Klug 2002, Kääb 2002, with more information). Ten bands of an ASTER scene from 29th May 2001 (Fig. 1) were selected for classifi cation, i.e. all the visible/near-infrared (VNIR) and short wave infrared (SWIR) bands together with thermal infrared (TIR) band 11. The bands were stacked, geome-trically rectifi ed to fi xed landmarks in a monochrome or-thophoto of 1996 and simultaneously resampled by cubic convolution to a resolution of 5 m to match DEM resoluti-on. As in other monotemporal investigations, atmosphe-ric and topographic corrections were rejected, since they can introduce additional errors into the data set (relevant reference in Schneevoigt et al. 2008). Several ratios were used in this work; the Normalised Difference Vegetation Index (NDVI) forms important thresholds in the classifi ca-tion hierarchy (Fig. 3).A digital elevation model (DEM) of 5 m ground resolution (Fig. 2 top) was generated and hydrologically stream cor-rected by SEDAG partners using photogrammetric data by the Bavarian Geodetic Survey. It served for the generation of fi ve DEM derivatives. These geomorphometric grids of horizontal, vertical and total curvature, slope and aspect were incorporated in the classifi cation process in addition to the DEM (Fig. 2). The landforms considered in this study are summarized in Tab. 1 (additional information can be found in Schnee-voigt et al. 2008). They result from interacting and parti-ally equifi nal processes. Thus strict delimitations of land-forms do not always exist in landscape: many forms show no clear boundaries (Figs. 1, 2). As partially interfi ngered deposits are frequent, form characteristics deviate from the ideals. This ”fuzzy nature of most high-mountain ter-rain features” (Kääb 2002) makes it necessary to consider context for sound classifi cations. Keeping the interval of

55

raum Nora J. Schneevoigt et al.

time in mind, the orthophoto of 1996 served as ground truth for crest and rockwall regions as well as for the sedi-ment stores on the valley fl oor. Digital photos taken, as far as accessible, from opposed slopes and peaks, were used

for the same purpose. A geomorphological map drawn up by Schrott et al. (2003) served as reference when dealing with the Reintal valley bottom.

Figure 1 : A: Location of the Reintal in the Northern Calcareous Alps. B: The valley stretches in an east-westerly direction along the Austro-German border, which runs on the southern valley crest (ASTER drape over DEM).

upper regions rockwalls valley bottom ubiquitary

snow and ice grass covered slopes shrub covered talus bare rock (<50°)

eastern cirque wall (<50°) shrub covered slopes tree covered talus bare rock (>50°)

eastern cirque wall (>50°) tree covered slopes alluvial fan fi ne sediments

western cirque wall (<50°) fl oodplain coarse sediments

western cirque wall (>50°) rockfall deposits channel

vegetation covered channel

debris/grass covered talus

Table 1: Target classes in the classifi cation process, sorted according to their predominant location in the study area.

56


4. Image segmentation into objects

The object-oriented approach (not to be confused with the homonymous programming mode) consists of two separate steps. First, the data employed is segmented into homogeneous image objects through generalisation and average determination, smoothing out irregular pixel-dominated patterns and creating more realistic forms. Secondly, these entire objects are classifi ed, not individual

pixels. Object-oriented image analysis hence unites the spectral analyses of remote sensing and the geometric tools of GIS into one desktop environment. The segmentation algorithm by Baatz and Schäpe (2000) segments an image in a knowledge-free way via region-growing, an automatized heuristic optimization method: the potential increase of spectral heterogeneity is assessed in a merge weighed by the size of two pixels or segments considered. Next to this colour criterion based on spec-

Figure 2: Top: Data basis employed in this study. Bottom: Segmentations of the four levels of the main project for landform detection.

57


tral information alone, shape parameters can be used to correct highly textured data which otherwise would pro-duce frayed and distorted segments. This constitutes an advantage especially in high mountain data. Yet it must be applied carefully, as it implies an arbitrary divergence from the given spectral information based on pure arith-metics (Baatz and Schäpe 2000). Not only spectral data, but also DEMs and all kinds of derivatives from image and elevation information can be integrated in the segmen-tation process. The available sources or parameters can be weighed by factors to differentiate their respective in-fl uence on the creation of a layer. These assets compen-sate for the extra time-consumption of fi nding adequate segmentation parameters. To address features at diffe-rent scales, individual layers must be segmented for each scale. The multiresolution segmentation algorithm (Baatz and Schäpe, 2000) permits a simultaneous depiction of several image levels segmented at various spatial resolu-tions (Benz et al. 2004). Yet the integration of additional levels only makes sense if this implies a gain of informati-on which cannot be retrieved from the existing levels. Initially, a strata mask distinguishing three altitudinal storeys was generated in an ancillary project with three levels. In this set of preclassifi cations, crest regions and valley bottom both carry an error of commission to gua-rantee inclusion of all relevant image objects. The resul-

ting layer was imported as L4 into the main project (Fig. 2). Here, segmentation on four levels was necessary in order to create the boundaries for all target classes. A small sca-le parameter conveys the spectral ground information of the Reintal (L1, Fig. 2). Conversely, the imported mask of three altitudinal subsystems (L4) requires a very high scale parameter, and the cirques and hanging valleys (L3) a re-latively high one. An intermediate level serves for the fi nal classifi cation (L2), so that smaller landforms can be displa-yed while preserving a certain degree of generalization. The segmentation of level L2 comprises the scales and parameters necessary to optimally depict the different landforms. For example, decreasing the colour criterion improves the representation of water bodies, but worsens the taluses at the same time. VNIR bands and DEM deri-vatives were brought into an equilibrium of 4:3, so that spectral information dominates. Scale 13 guaranteed the existence of necessary boundaries; decisive further ame-liorations only set in below scale 10. However, this would have increased the project and processing times on the one hand, while leading to a lesser degree of abstraction due to small image objects on the other. Hence paramete-risation resulted in level L2 illustrated in Fig. 2.

Figure 3: Extract of the classifi cation hierarchy of the main project. The fi gures around the membership function icon to the right indicate the bounds of the fuzzy areas.

58


5. Classification of the image objects

Each image object can be addressed by mean value, stan-dard deviation and ratio of the incorporated pixels next to its individual geometric features and its neighbourhood. When working on several image levels, relationships, sub- and superordinations, morphometric and class-related features can also be used for class descriptions (Benz et al. 2004, Blaschke et al. 2002). This accommodates three-dimensional alpine applications, provided that segmenta-tion leads to objects which describe natural features geo-

metrically well (Schneevoigt et al. 2008).The four segmented levels were classifi ed individually af-ter developing the corresponding classifi cation hierarchy (Fig. 3). This knowledge base is edited from class descrip-tions, which divide into contained features immanent in a class itself and inherited ones passed on by parent classes in the class hierarchy. A combination of hard (L1) and soft (L2, L3, L4) classifi ers enables this approach generally re-sting on fuzzy logic. Hereby, fl oating thresholds provide a margin for the attribution of an object to a class (Fig.

3, right). Soft membership classifi ers return fuzzy values between 0 (no assignment at all) and 1 (full assignment) for each feature and image object considered. Besides, fuzzy logic operators, which produce for instance sums, subsets and means, link different feature terms (Baatz and Schäpe 2000).

6. Results

The classifi cation of level L1 renders ground land cover, level L4 the strata mask, level L3 eastern and western walls of cirques and hanging valleys (see Schneevoigt and

Figure 4: The fi nal landform classifi cation on level L2. Its 20 distinct classes completely cover the study area, leading to a coherent thema-tic map which approximates ground truth well.

Schrott 2006). This leads to a sound L2 landform classi-fi cation (Fig. 4). The majority of classes such as cirques, rockwalls, fl oodplains and sediments are identifi ed well. Detection limits are reached with moraine and rockfall deposits, because they have been overprinted by more recent processes for centuries or millennia and therefore leave no characteristic marks on the land surface. Then again, some target classes are further differentiated than previously expected. For instance, the vegetation cover of slopes and taluses was subdivided into high, medium and low natural cover, leading to 20 thematic landform classes (Tab. 1) in Fig. 4.

59


Table 2: Confusion table of the L2 classifi cation.

The fi nal landform classifi cation scores high in the assess-ments of both overall accuracy (92%), kappa coeffi cient (0.915), user’s and producer’s accuracy (Tab. 2). Only a few misclassifi cations occur, but they concern high amounts of pixels, as level L2 consists of image objects of ten to hundreds of pixels. Fuzzy classifi cation stability, i.e. the de-gree of distinctness between most and second most pro-bable class affi liation, is lower (Fig. 5), but best member-ship assignments score generally high, too. Alluvial fans tend to intermingle with fl oodplains, while talus sheets and cones could not be differentiated from one another. This owes to the fact that in situ, these landforms tend to mostly coalesce, so that their exact assignment relies on interpretation by the observer. Overall, the vegetation covers of talus show the highest confusion.

7. Discussion

Image segmentation represents an additional, time con-suming step in the classifi cation routine. Finding an opti-mal segmentation for L2 constituted a veritable challen-ge, because the often very faint signals from avalanche and debris fl ow tracks should still be captured without ending at too small a resolution. VNIR bands have to be given considerable weight, as they trace landforms best. Yet the object-oriented approach makes the diffi cult high

mountain terrain manageable (Schneevoigt et al. 2006) and leads to sound results. It remains to be investigated to which extend a purely pixel-based classifi cation scheme may handle this data.The good results partially owe to the fact that two distinct data sources were combined for analysis: some target classes appear spectrally distinct (e.g. sediments, rocks vs. vegetation covered features), whereas other landforms could be separated by topographic information. Giles and Franklin (1998) investigating geomorphological slope units reach an overall accuracy of 88.5% in their supervi-sed classifi cation based on prior image segmentation and classifi cation from training areas. However, the values in the confusion matrix (Tab. 2) have to be taken with care: on the one hand, object-oriented accuracy assessments tend to overestimate as they are based on averaged image objects and not on individual pixels. On the other, test areas were selected randomly, but without following a re-gular spatial pattern. Hence further accuracy assessment with different, pixel-based software is required. Besides, the classifi cation quality of older rockfall and moraine de-posits could not be assessed, as they form no classes in the hierarchy.Varying illumination constitutes a problem in high moun-tain areas. It can partially be mended by band ratio for-

60


Figure 5: Fuzzy classifi cation stability of level L2. Red = close proximity of best and second best membership assignment; green = distal, stable assignments. The north-easternmost part of the valley shows the most unstable memberships.

mation, i.e. the division of adjacent satellite image bands: discrepancies between them are reinforced, while similar structures are simultaneously eliminated. Hence while useful features like outlines of vegetation or ice/snow appear more clearly, atmosphere and relief induced va-riations in illumination disappear, as they are highly cor-related in neighbouring bands (Paul 2000). Several ratios were used in this work; the Normalised Difference Vegeta-tion Index (NDVI) forms important thresholds in the clas-sifi cation hierarchy (Fig. 3).

8. Conclusion and outlook

To further evaluate the results, the exact infl uences of image and DEM data respectively should be assessed by analysing them individually. Moreover, the transferability both of the segmentation parameterisation and the clas-sifi cation hierarchy still has to be investigated. One can assume that the application of such a two-step routine poses double problems. Conversely, Blaschke et al. (2002) argue that object-oriented classifi cation rules should be easier transferable than pixel-based ones, as the former depend less on refl ection values and atmospherical con-ditions. When transferring the methodology developed in this study both to other datasets and regions, the appro-priateness of NDVI application should also be compared

to soil-adjusted vegetation indices (for details see Schnee-voigt et al. 2008). Many open questions remain to be answered in this in-terdisciplinary work linking geomorphology and remo-te sensing. Albertz (2001) stresses that the appropriate analysis of remotely sensed imagery can become highly diffi cult when operating between disciplines, as remote sensing methods are not delivered with problem-adapted assessment factors. Then again, a broadened knowledge on sediment storage features represents the prerequisite for further insights into processual behaviour and land-form development in the fragile mountain environment (Schrott et al. 2003). Considering the good match of the fi nal landform classifi cation and ground truth, the object-oriented approach constitutes a valuable tool for the Al-pine sediment cascade, especially in inaccessible regions.

Acknowledgements

Many thanks go to the Remote Sensing Laboratories, Uni-versity of Zürich (K. Itten), the Center for Remote Sensing of Land Surfaces (M. Braun) and the Remote Sensing Re-search Group (G. Menz, H.-P. Thamm), both University of Bonn, for providing workplace and software. The DEM was generated by SEDAG partners using photogrammetric data by the Bavarian Geodetic Survey. The ASTER scene

61


References

ALBERTZ, J., 2001: Einführung in die Fernerkundung. Grundlagen der Interpretation von Luft- und Satelliten-bildern. Wissenschaftliche Buchgesellschaft, Darmstadt.

BAATZ, M., SCHÄPE, A., 2000: Multiresolution segmenta-tion - an optimization approach for high quality multi-scale image segmentation. In: Strobl, T., Blaschke, T., Griesebner, G. (eds.), Angewandte Geographische Infor-mationsverarbeitung XII, Beiträge zum AGIT-Symposium Salzburg 2000. Herbert Wichmann Verlag, Heidelberg, 12–23.

BENZ, U.C., HOFMANN, P., WILLHAUCK, G., LINGENFELDER, I., HEYNEN, M., 2004: Multiresolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready in-formation. ISPRS Journal of Photogrammetry & Remote Sensing, 58, 239–258.

BLASCHKE, T., GLÄSSLER, C., LANG, S., 2000: Bidverarbei-tung in einer integrierten GIS/Fernerkundungsumge-bung - Trends und Konsequenzen. In: T. Blaschke (ed.): Fernerkundung und GIS. Neue Sensoren - innovative Me-thoden. Herbert Wichmann Verlag, Heidelberg, 1–9.

CAINE, N.T., 1974: The geomorphic processes of the alpine environment. In: Ives, J., Barry, R. (eds.): Arctic and Alpine Environments. Methuen, London, 721–748.

GILES, P.T., FRANKLIN, S.E., 1998: An automated approach to the classifi cation of the slope units using digital data. Geomorphology, 21(3-4), 251–264.

KÄÄB, A., 2002: Monitoring high-mountain terrain defor-mation from repeated air- and spaceborne optical data: examples using digital aerial imagery and ASTER data. ISPRS Journal of Photogrammetry & Remote Sensing, 57, 39–52.

KLUG, H., 2002: Tutorial Version 1.1: Eine Einführung in die Verwendung von ASTER. Asterdaten und ihre Ver-wendung im landschaftsökologischen Kontext. LARG Technical Report, Universität Salzburg, http://www.geo.sbg.ac.at/larg/Astertutorial.pdf, 15.08.2006.

PAUL, F., 2000: Evaluation of different methods for gla-

cier mapping using Landsat-TM data. In: Wunderle, S. (ed.): EARSeL Workshop on Remote Sensing of Land Ice and Snow, Juni 2000, Dresden, 239–245.

SCHNEEVOIGT, N.J., VAN DER LINDEN, S., THAMM, H.-P. & L. SCHROTT (2008): Detecting Alpine landforms from remotely sensed imagery. A pilot study in the Bavarian Alps. Geomorphology, 93: 104-119.

SCHNEEVOIGT, N.J., SCHROTT, L., 2006: Linking geomor-phic systems theory and remote sensing. A conceptual approach to Alpine landform detection (Reintal, Bavarian Alps, Germany). Geographica Helvetica 3, 181-190.

SCHROTT, L., HUFSCHMIDT, G., HANKAMMER, M., HOFF-MANN, T., DIKAU, R., 2003: Spatial distribution of sedi-ment storage types and quantifi cation of valley fi ll depo-sits in an alpine basin, Reintal, Bavarian Alps, Germany. Geomorphology, 55, 45 – 63.

was provided by EOS Data Gateway (NASA) and is cour-tesy of NASA/GSFC/METI/ERSDAC/JAROS and the US/Ja-pan ASTER science team. This study was embedded in the SEDAG projects, which were funded by the German Re-search Foundation (DFG) from 2000 to 2008. N.J. Schnee-voigt was supported by the German National Academic Foundation until 2004, then by SEDAG and in 2006 within the International Scholarship Programme of the Gottlieb Daimler and Karl Benz Foundation. Sebastian van der Lin-den was funded by the scholarship programme of the Ger-man Federal Environmental Foundation (DBU) until 2007.

62


Correspondence to:

NORA JENNIFER SCHNEEVOIGT

Department of Geosciences

University of Oslo

Postbox 1047 Blindern, N-0316 Oslo, Norway

e-mail: [email protected]

SEBASTIAN VAN DER LINDEN

Geomatics Lab

Humboldt-University Berlin

Rudower Chaussee 16, 12489 Berlin


TOBIAS KELLENBERGER

Topografie

swisstopo

Seftigenstr. 264, CH-3084 Wabern, Switzerland


ANDREAS KÄÄB

Department of Geosciences

University of Oslo

Postbox 1047 Blindern, N-0316 Oslo, Norway


LOTHAR SCHROTT

Department of Geography and Geology

University of Salzburg

Hellbrunnerstr. 34, A-5020 Salzburg, Austria


object-oriented classification of alpine landforms from an ... · a digital elevation model (dem)...

Documents